Execution

execution controls how the run behaves: the order attacks run in, how much runs at once, what happens after an error, and where results go. Every field has a safe default, so you can leave the whole section out.

execution

How the run behaves: ordering, limits, error handling, output and storage. Every field has a safe default.

Field	Type	Default	Description
mode	one of sequential, parallel	sequential	Run attacks one after another, or all at once within the shared limits.
dry_run	boolean	false	Validate and build everything, but send no request.
preflight	boolean	true	Check every model the campaign uses (target, guardrails, attack roles and embedders, judges, goal classifier) before attacking, so an unreachable endpoint fails fast with a clear message instead of mid-run. Turn it off to let a run proceed and record such failures as attempt errors.
on_error	one of continue, stop	continue	After an attack fails, run the remaining ones or stop.
escalate	boolean	false	Run attacks in order and drop a goal once an attack jailbreaks it, so later attacks only face the goals still standing. Implies sequential order; the remaining attacks are skipped once every goal has fallen.
per_attack_timeout	number	unset	Maximum seconds a single attack may run before it is stopped and marked failed. Leave unset for no limit.
concurrency	ConcurrencySpec	see ConcurrencySpec	How much work may run in parallel.
retries	RetrySpec	see RetrySpec	Extra attempts after a provider error.
output	OutputSpec	see OutputSpec	Files written for the run.
storage	StorageSpec	see StorageSpec	Where results are recorded.

execution.concurrency

How much work may be in flight at once, shared by every attack.

Field	Type	Default	Description
attack	integer (>= 1)	1	How many goals may be generating requests at the same time.
target	integer (>= 1)	1	How many requests may be waiting on the target at the same time.
judge	integer (>= 1)	1	How many replies may be judged at the same time.

execution.retries

Extra attempts after a provider error (timeouts, rate limits, 5xx).

Field	Type	Default	Description
target	integer (>= 0)	0	Extra attempts for a failed call to the target.
roles	integer (>= 0)	0	Extra attempts for a failed call to an attack's helper models.
judges	integer (>= 0)	0	Extra attempts for a failed call to a judge.

execution.output

Files written for each run, next to the results stored in the database.

Field	Type	Default	Description
directory	string	./logs/runs	Folder where runs are written, each in a subfolder named after its run id.
run_id	string	unset	Fixed id for the run. Leave unset to generate a new one each time.
formats	list of one of json, jsonl	["json"]	File formats to write: json (one document) and/or jsonl (one line per attempt).
save_prompts	boolean	true	Include the prompts sent to the target.
save_responses	boolean	true	Include the target's replies.
save_traces	boolean	true	Include the search traces of adaptive attacks (every call they made).

execution.storage

Where results are recorded so they can be browsed later.

Field	Type	Default	Description
backend	one of local, remote	local	local writes to a SQLite file on this machine; remote sends results to the HackAgent server. The backend is always chosen here, never inferred from which API keys happen to be set, and remote fails before the run starts if its key is missing.
base_url	string	unset	Address of the remote server. Defaults to HACKAGENT_BASE_URL or the hosted service.
api_key	string	unset	Runtime API key, defaulting to HACKAGENT_API_KEY. In a file use an explicit reference such as ${HACKAGENT_API_KEY}.

Examples

Fail fast on an unreachable endpoint, then run several goals at once

execution:
  preflight: true
  per_attack_timeout: 600
  concurrency:
    attack: 4
    target: 8
    judge: 4

Escalate: stop spending attacks on goals that already fell

execution:
  escalate: true